Ok I see now that the "compute" variable in the code and the training run sizes take into account effective compute, which IMO seems confusing. Anyway, at least in the playground IMO the "biggest training run" seems misleading as it sounds like physical compute
---- Previously said -----
I noticed an issue when playing with https://takeoffspeeds.com/ and realized that for some reason the full automation training run was always at least as big as the 2022 AGI training requirements.
Looking at the code, I realized that the automation seems to be done using physical compute (biggest training run) with no adjustment for better software. Seems like a pretty important error to fix if so?
This seems to make more sense, and also match the report https://docs.google.com/document/d/1rw1pTbLi2brrEP0DcsZMAVhlKp6TKGKNUSFRkkdP_hs/edit#heading=h.qh3st3bob46f
Ok I see now that the "compute" variable in the code and the training run sizes take into account effective compute, which IMO seems confusing. Anyway, at least in the playground IMO the "biggest training run" seems misleading as it sounds like physical compute
---- Previously said -----
I noticed an issue when playing with https://takeoffspeeds.com/ and realized that for some reason the full automation training run was always at least as big as the 2022 AGI training requirements.
Looking at the code, I realized that the automation seems to be done using physical compute (biggest training run) with no adjustment for better software. Seems like a pretty important error to fix if so?
This seems to make more sense, and also match the report https://docs.google.com/document/d/1rw1pTbLi2brrEP0DcsZMAVhlKp6TKGKNUSFRkkdP_hs/edit#heading=h.qh3st3bob46f